arunp0/yama-indic

Yama: zero-shot text classifier for Indian languages and Hinglish

Yama (yama-indic) is an open-weights, Apache-2.0 zero-shot text classifier for Hindi, Hinglish, 10 other Indian languages and English. You pass a text and plain-English questions with your own labels; it returns a calibrated probability for every label in one forward pass, without fine-tuning. It is a 586M-parameter encoder (XLM-RoBERTa-large), not a generative LLM, so answers are fast, deterministic and cheap to run.

On tasks Yama never trained on (33 tasks, 19 languages)	Accuracy
Yama v3	0.814
MoritzLaurer/bge-m3-zeroshot-v2.0	0.770
joeddav/xlm-roberta-large-xnli	0.560
MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7	0.478
facebook/bart-large-mnli	0.374

Same test samples and harness for every model; details under Comparison with other zero-shot classifiers.

Use it for: intent detection and routing, support-ticket triage, sentiment, topic and news classification, yes/no checks ("is this a complaint?") and 1–N ratings ("how urgent?"), on text in Hindi (Devanagari or romanised), Hinglish, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Urdu, Odia and English. Labels, questions and the number of options are defined at inference time, so one model handles many tasks.

How it works

Yama is a typed-decision model. You give it a text (a message, ticket, review, headline or document) and one or more questions written in plain English. It answers every question in a single forward pass, with a calibrated probability for each option. Question schemas are defined at inference time, so it handles new labels and new tasks without retraining.

The name. In Indian tradition Yama is also Dharmaraja, the judge who weighs what is placed before him and gives a verdict. This model does the same with the options you give it.

Formerly laya-indic. This repository was renamed with v3; old links to arunp0/laya-indic redirect here. The v1 and v2 weights are kept at revisions v1.0 and v2.0.

Yama is an independent personal release. It is not affiliated with or endorsed by HAL-X AI. It uses the THX-01 decision architecture and the open-source thx01 library (Apache-2.0, HAL-X AI), but it is built on the FacebookAI/xlm-roberta-large encoder (MIT), not on THX-01's weights. All training data is public and permissively licensed, plus labels generated and verified by open-weights models (Qwen3, Qwen3.6, Gemma 4). No proprietary-model output was used anywhere.

Parameters	about 586M (XLM-R-large encoder plus a 2-layer decision head)
Input	up to 512 tokens per question (question, options and text)
Question types	choice, noul (yes/no), score (ordinal)
Training	3 epochs over 591,217 decisions: 285,078 from 17 public datasets, 173,226 verified Qwen3 synthetic questions, 117,646 verified questions about real Indian-language text and 15,267 verified cause/effect and passage yes/no questions
Languages	trained on 30+ languages; strongest on English, Hindi, Hinglish and 10 other Indian languages
Licence	Apache-2.0

What's new in v3

• A larger encoder. v3 is built on XLM-R-large (560M) instead of mmBERT-base. Overall accuracy rises from 0.800 to 0.862 and mean calibration error falls from 0.091 to 0.070. The biggest gains are in Indian languages: SIB-200 Indic topic 0.670 → 0.844, IndicSentiment 0.884 → 0.950, MASSIVE Indic 0.849 → 0.893.
• Causal reasoning is no longer at chance. IndicCOPA rises from 0.513 to 0.624. Telugu and Gujarati, which had no cause/effect training data at all, improved as much as the other languages (0.528 → 0.634, 0.529 → 0.632).
• More real Indian-language text. 54,198 new questions about real Sangraha text, written by Gemma 4 and verified by Qwen3.6, with extra weight on Odia and Urdu, which v2 barely covered.
• Trade-offs. About twice the compute per decision of v2, and the input limit is 512 tokens instead of 1,024. Banking77 (0.819 → 0.791) and SMS spam (0.964 → 0.944) are lower than v2. BoolQ-hi is unchanged (0.618 → 0.616).

Question types

type	returns	criteria
choice	one of N options, with a probability for each	{"key": "description", ...} (descriptions may be null)
noul	P(yes)	none
score	an ordinal level and its distribution	an ordered list of level descriptions

THX-01 also supports number, excerpt and "cite": true. This model was not trained for those and does not support them.

Quickstart

import thx01   # pip install thx01

agent = thx01.load("arunp0/yama-indic")          # or a local directory; GPU if available
# earlier weights: thx01.load(huggingface_hub.snapshot_download("arunp0/yama-indic", revision="v2.0"))

text = "Kal se mera UPI payment fail ho raha hai, paise kat gaye par merchant ko nahi mile"
out = agent.system_one(text, {
    "topic":  {"type": "choice", "instructions": "What is the customer writing about?",
               "criteria": {"failed_payment": "payment failed or money debited but not received",
                            "refund": "asking for a refund", "account": "login or account access", "other": None}},
    "urgent": {"type": "noul", "instructions": "Does the customer need help urgently?"},
    "anger":  {"type": "score", "instructions": "How frustrated is the customer?",
               "criteria": ["calm", "mildly annoyed", "clearly frustrated", "furious"]},
})
# topic  -> failed_payment (0.96)
# urgent -> P(yes) 0.74
# anger  -> "clearly frustrated" 0.88

The response's model field says "THX-01". That is the thx01 library's default label; the weights are this model's.

For many texts, use agent.predict_batch(texts, questions, batch_size=64). questions can be one schema shared by every text or a list with one schema per text.

Write questions in English and leave the text in its own language and script. That is how the model was trained.

Benchmarks

Zero-shot accuracy on 67 tasks, compared with THX-01 (322M, same architecture, roughly 2.1M training decisions) and with v2 (laya-indic v2). All models were scored with the same harness, the same prompts and the same test samples. The harness is adapted from THX-01's open benchmark code (bench/tasks.py).

Task family	Tasks	THX-01	v2	v3
Overall (all tasks)	67	0.758	0.800	0.862
English general (AG News, Emotion, Banking77, SMS, MASSIVE en)	7	0.814	0.851	0.853
MASSIVE scenario, THX's 14 languages	14	0.871	0.878	0.901
MASSIVE intent@20 + yes/no (az, hi)	4	0.916	0.919	0.931
SIB-200 topic, THX's 10 languages	10	0.768	0.779	0.844
SIB-200 topic, Indic (hi + 9)	10	0.630	0.670	0.844
MASSIVE scenario, Indic (bn ta te kn ml ur)	6	0.669	0.849	0.893
IndicSentiment (6 languages)	6	0.878	0.884	0.950
IndicCOPA (6 languages)	6	0.524	0.513	0.624
Hinglish-TOP (domain, intent@20)	2	0.789	0.940	0.942
Synthetic held-out schemas (Qwen3)	4	0.718	0.872	0.924

Per task: v3 is ahead of THX-01 on 63 of 67 tasks, behind on 3, tied on 1, and ahead of v2 on 60, behind on 6, tied on 1.

Comparison with other zero-shot classifiers

Four widely used zero-shot classifiers (the Hub's most-downloaded one, bart-large-mnli, which is English-only, and three multilingual ones), scored on the same test samples with the same harness. Only choice tasks are compared (60 tasks), because NLI models have no native yes/no-question or rating mode. For the NLI models each option becomes the hypothesis "This text is about {option description}." (COPA alternatives are used as-is), and probabilities are a softmax over entailment scores, as in the Hugging Face zero-shot-classification pipeline.

Task family	Tasks	Yama v3	bart-large-mnli	mDeBERTa-v3-base-xnli	xlm-roberta-large-xnli	bge-m3-zeroshot-v2.0
Not in Yama's training data	33	0.814	0.374	0.478	0.560	0.770
SIB-200 topic (19 languages)	19	0.843	0.273	0.321	0.504	0.788
IndicSentiment (6 languages)	6	0.950	0.502	0.857	0.768	0.888
IndicCOPA (6 languages)	6	0.624	0.479	0.575	0.553	0.586
AG News	1	0.878	0.785	0.774	0.709	0.910 †
DAIR Emotion	1	0.535	0.491	0.316	0.279	0.693 †
Indian-language subset of the above	22	0.813	0.347	0.532	0.586	0.753
Train split in Yama's data (not zero-shot for Yama)	27	0.902	0.200	0.304	0.405	0.669

Full model ids: facebook/bart-large-mnli, MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7, joeddav/xlm-roberta-large-xnli, MoritzLaurer/bge-m3-zeroshot-v2.0.

• On the 33 tasks not in its training data, Yama is ahead of bart-large-mnli, mDeBERTa-xnli and xlm-roberta-large-xnli on every task, and ahead of bge-m3-zeroshot-v2.0 on 29, behind on 3 (AG News, DAIR Emotion, IndicCOPA Marathi) and tied on 1. Mean calibration error on these tasks is 0.092 for Yama and for bge-m3-zeroshot-v2.0, 0.100 for bart-large-mnli, 0.140 for xlm-roberta-large-xnli and 0.223 for mDeBERTa-xnli.
• The bottom row (MASSIVE, Banking77, Hinglish-TOP) is shown for completeness only: Yama was trained on those datasets' training splits and the NLI models were not.
• Sanity check: our harness gives the NLI models scores at or above their own published results (for example bart-large-mnli on AG News: 0.785 accuracy here, 0.703 macro-F1 on the bge-m3-zeroshot card), so the prompt wording is not handicapping them.
• † The bge-m3-zeroshot-v2.0 card reports for AG News and DAIR Emotion only the score of a run trained on up to 500 examples per class of those datasets (0.892 and 0.697 F1), and our scores match those, so its AG News and Emotion results are probably not zero-shot.
• Yama's own synthetic training data includes generic topic, sentiment and emotion questions (not these datasets), so "not in training data" means the dataset, not the task type.

Read these numbers with the training data in mind. None of the models saw these test sets. Some tasks, however, have their training splits or similar task formats in one model's data and not in the other's:

Eval family	This model trained on	THX-01
MASSIVE	train split, 20 languages (incl. hi bn ta te kn ml ur, az ka tr fa)	train splits ("public benchmarks' training splits")
Hinglish-TOP	train + validation splits	not stated
Banking77, SMS spam	Banking77 train; SMS messages outside the 1,000-message test sample	not stated
AG News	nothing (licence unclear, so excluded)	not stated individually
SIB-200	nothing (CC-BY-SA, so excluded); synthetic topic sentences and Qwen3/Gemma-labelled topics on real Sangraha text	not stated
IndicCOPA	no IndicCOPA data. New in v3: 7,782 cause/effect questions in the COPA format, written by Gemma 4 from real Sangraha text in 9 languages. Telugu and Gujarati were deliberately left out to measure transfer	not stated
BoolQ-hi	no BoolQ data. New in v3: 7,485 passage yes/no questions on real Sangraha text in 6 languages. Hindi was deliberately left out	not stated
IndicSentiment, DAIR Emotion	nothing; synthetic reviews and emotion datasets instead	not stated

All training texts, including the Sangraha corpus the real-text questions were cut from (line by line), were deduplicated against every evaluation set, using normalised exact match.

Per-task results (accuracy and expected calibration error)

Task	n	THX-01 acc	v2 acc	v3 acc	THX-01 ECE	v3 ECE
AG News (4)	1000	0.931	0.864	0.878	0.009	0.043
DAIR Emotion (6)	1025	0.427	0.492	0.535	0.261	0.284
Banking77 (77)	1001	0.658	0.819	0.791	0.047	0.086
SMS spam (yes/no)	1000	0.883	0.964	0.944	0.101	0.035
MASSIVE-intent@20 (en)	1000	0.929	0.925	0.933	0.065	0.052
Intent yes/no (en)	1000	0.961	0.967	0.962	0.035	0.030
MASSIVE-intent@20 (az)	1000	0.909	0.881	0.905	0.051	0.059
Intent yes/no (az)	1000	0.948	0.945	0.948	0.028	0.025
MASSIVE-scenario (en)	500	0.906	0.926	0.930	0.089	0.054
MASSIVE-scenario (az)	500	0.878	0.862	0.900	0.061	0.051
MASSIVE-scenario (ru)	500	0.896	0.900	0.908	0.067	0.045
MASSIVE-scenario (tr)	500	0.874	0.866	0.886	0.084	0.067
MASSIVE-scenario (de)	500	0.884	0.902	0.918	0.068	0.048
MASSIVE-scenario (fr)	500	0.894	0.902	0.912	0.066	0.042
MASSIVE-scenario (es)	500	0.868	0.902	0.892	0.070	0.056
MASSIVE-scenario (ar)	500	0.830	0.844	0.838	0.067	0.053
MASSIVE-scenario (hi)	500	0.856	0.892	0.900	0.051	0.050
MASSIVE-scenario (zh-CN)	500	0.872	0.898	0.910	0.072	0.057
MASSIVE-scenario (fa)	500	0.894	0.890	0.920	0.075	0.053
MASSIVE-scenario (ka)	500	0.762	0.728	0.870	0.055	0.064
MASSIVE-scenario (ja)	500	0.908	0.922	0.926	0.066	0.052
MASSIVE-scenario (ko)	500	0.868	0.856	0.910	0.053	0.045
SIB-200 (az)	204	0.779	0.770	0.853	0.091	0.062
SIB-200 (en)	204	0.833	0.819	0.828	0.068	0.077
SIB-200 (ru)	204	0.779	0.848	0.848	0.076	0.058
SIB-200 (tr)	204	0.779	0.745	0.853	0.101	0.052
SIB-200 (de)	204	0.819	0.819	0.848	0.086	0.070
SIB-200 (ar)	204	0.774	0.779	0.828	0.069	0.068
SIB-200 (hi)	204	0.721	0.730	0.868	0.132	0.047
SIB-200 (zh)	204	0.794	0.828	0.843	0.066	0.072
SIB-200 (kk)	204	0.701	0.760	0.843	0.126	0.062
SIB-200 (uz)	204	0.701	0.691	0.824	0.127	0.083
Hinglish-TOP domain	1000	0.887	0.969	0.976	0.057	0.043
Hinglish-TOP intent@20	1000	0.691	0.910	0.907	0.050	0.058
MASSIVE-intent@20 (hi)	1000	0.877	0.904	0.911	0.033	0.044
Intent yes/no (hi)	1000	0.932	0.945	0.960	0.028	0.021
BoolQ-hi (yes/no) ‡	1000	0.647	0.618	0.616	0.186	0.236
MASSIVE-scenario (bn)	500	0.698	0.846	0.884	0.077	0.045
MASSIVE-scenario (ta)	500	0.700	0.848	0.870	0.077	0.043
MASSIVE-scenario (te)	500	0.622	0.838	0.900	0.082	0.050
MASSIVE-scenario (kn)	500	0.614	0.854	0.906	0.066	0.051
MASSIVE-scenario (ml)	500	0.652	0.846	0.908	0.056	0.077
MASSIVE-scenario (ur)	500	0.728	0.862	0.892	0.059	0.055
SIB-200 (mr)	204	0.647	0.726	0.863	0.180	0.046
SIB-200 (bn)	204	0.623	0.662	0.819	0.168	0.061
SIB-200 (ta)	204	0.691	0.696	0.833	0.111	0.060
SIB-200 (te)	204	0.632	0.657	0.858	0.131	0.042
SIB-200 (gu)	204	0.618	0.672	0.833	0.135	0.065
SIB-200 (kn)	204	0.608	0.662	0.843	0.181	0.050
SIB-200 (ml)	204	0.529	0.598	0.853	0.248	0.073
SIB-200 (pa)	204	0.564	0.603	0.819	0.171	0.062
SIB-200 (ur)	204	0.672	0.691	0.853	0.138	0.055
IndicSentiment (hi)	998	0.916	0.927	0.957	0.099	0.019
IndicSentiment (bn)	998	0.868	0.878	0.949	0.065	0.022
IndicSentiment (mr)	998	0.894	0.908	0.949	0.080	0.017
IndicSentiment (ta)	998	0.869	0.847	0.951	0.063	0.019
IndicSentiment (te)	998	0.859	0.868	0.943	0.059	0.021
IndicSentiment (gu)	998	0.861	0.877	0.953	0.060	0.016
IndicCOPA (hi)	449	0.550	0.501	0.617	0.168	0.249
IndicCOPA (mr)	449	0.510	0.514	0.604	0.184	0.258
IndicCOPA (bn)	500	0.522	0.518	0.626	0.186	0.233
IndicCOPA (ta)	500	0.540	0.490	0.634	0.153	0.228
IndicCOPA (te)	500	0.532	0.528	0.634	0.175	0.232
IndicCOPA (gu)	448	0.491	0.529	0.632	0.219	0.216
Qwen3-synth held-out schemas (choice)	1008	0.726	0.845	0.911	0.119	0.053
Qwen3-synth held-out (Hinglish)	719	0.729	0.890	0.940	0.128	0.035
Qwen3-synth held-out schemas (noul)	1342	0.752	0.930	0.961	0.116	0.034
Qwen3-synth held-out schemas (score)	650	0.663	0.823	0.886	0.109	0.045

ECE is expected calibration error with 15 equal-width bins; lower is better. Mean ECE over the 67 tasks is 0.070 for v3, 0.091 for v2 and 0.097 for THX-01.

‡ BoolQ-hi is left out of every mean, as in earlier versions. Our harness had a label-parsing bug in the first runs of this benchmark; all rows here were scored after the fix.

Where v3 is stronger than THX-01 and v2:

• Indian-language topic classification (SIB-200 Indic +0.21 over THX-01, +0.17 over v2) and sentiment (+0.07 over THX-01).
• Low-resource languages THX-01 also covers: Georgian, Kazakh, Uzbek, Azerbaijani, Turkish.
• Causal reasoning in Indian languages (IndicCOPA +0.10), including the two languages it had no data for.
• New decision schemas it has never seen (synthetic held-out +0.21 over THX-01).

Where it is weaker:

• AG News (0.878 vs 0.931 for THX-01). THX-01 reports training on public benchmarks' training splits; AG News is excluded here.
• Banking77 and SMS spam are a little lower than v2.
• BoolQ-hi (0.616) is below THX-01 (0.647); passage yes/no questions remain hard.
• IndicCOPA (0.624) is well above chance but far from strong.

Training

Method. RLCD (Reinforcement Learning for Calibrated Decisions), as introduced by THX-01 and implemented in thx01:

• Policy-gradient term: sampling from noisy logits, rewarded by a strictly proper scoring rule.
• Soft cross-entropy term: on label-smoothed targets.
• Calibration: a temperature for each question type, fitted after training on held-out data.

Decision training starts from FacebookAI/xlm-roberta-large with a freshly initialised decision head. (Continued pretraining on Sangraha, which helped v2, gave no measurable gain on XLM-R-large, so v3 does not use it.)

Epochs	2, then a third epoch at one fifth of the learning rate
Batch	16 × 2 gradient accumulation = 32 decisions per update; 36,950 updates per 2 epochs
Learning rates	encoder 1e-5, head 2e-4 (epoch 3: 2e-6 / 4e-5); AdamW, weight decay 0.01, warmup, cosine decay
Precision	bf16 autocast, with a loss-spike guard (roll back and halve the learning rate if the running loss jumps); it never fired
Hardware / time	one NVIDIA H100: about 2.6 hours for epochs 1–2 and 1.3 hours for epoch 3
Question mix	268k choice, 221k yes/no, 102k score

Public data (285,078 decisions; unchanged from v2; every source is CC0, CC-BY, Apache-2.0 or MIT; full attribution in LICENSES.md):

Source	Decisions	Licence	Used for
MoritzLaurer/synthetic_zeroshot_mixtral_v0.1	59,998	Apache-2.0	general zero-shot choice / yes-no over diverse tasks
MASSIVE (20 languages, train)	54,875	CC-BY-4.0	intent and domain choice, intent yes/no
HuffPost News Category	28,552	CC-BY-4.0	news-topic choice, coarse and fine
HelpSteer2	20,000	CC-BY-4.0	response-quality score questions
BRIGHTER emotion	19,344	CC-BY-4.0	emotion yes/no, human-labelled in-language (Hindi and Marathi in full, plus 8 other languages)
Civil Comments	12,999	CC0-1.0	toxicity yes/no and score
GoEmotions	11,998	Apache-2.0	fine-grained emotion
Nemotron Safety Guard v3	11,000	CC-BY-4.0	multilingual safety categories
Aegis 2.0	10,999	CC-BY-4.0	safety categories
LEDGAR (LexGLUE)	10,000	CC-BY-4.0	contract-clause type
Banking77 (train)	9,964	CC-BY-4.0	77-way intent
WMT DA human evaluation	9,000	Apache-2.0	translation-quality score
Measuring Hate Speech	8,000	CC-BY-4.0	hate-speech facets
IBM Argument Quality 30k	8,000	CC-BY-3.0	argument-quality score
Hinglish-TOP (train + validation)	6,936	Apache-2.0	Hinglish domain and intent
Aya Red-teaming	2,685	Apache-2.0	multilingual harm categories
SMS Spam Collection	728	CC-BY-4.0	spam vs ham

Model-generated data (306,139 verified questions). Every question was answered blind by a second, different open-weights model and kept only if the two agreed.

Part	Texts	Questions	Generator → verifier	New in
Label-first synthetic decisions	59,310	173,226	Qwen3-235B-A22B-Instruct-2507 → Qwen3-30B-A3B-Instruct-2507	v1
Questions on real Sangraha text	22,791	63,448	Qwen3-235B-A22B-Instruct-2507 → Qwen3-30B-A3B-Instruct-2507	v2
Questions on real Sangraha text	18,699	54,198	Gemma-4-26B-A4B-it → Qwen3.6-35B-A3B	v3
Cause/effect questions (COPA format) from real Sangraha text	7,782	7,782	Gemma-4-26B-A4B-it → Qwen3.6-35B-A3B	v3
Passage yes/no questions on real Sangraha text	3,192	7,485	Gemma-4-26B-A4B-it → Qwen3.6-35B-A3B	v3

All five models are open-weights and Apache-2.0. The v3 labels were generated on our own GPU with vLLM. The three v3 parts are published as arunp0/yama-indic-gemma (configs real_text, copa, boolq, with held-out test splits).

• Synthetic decisions (v1): the generator proposes a schema, gold answers are sampled first, then it writes a text that makes them true, in a chosen language, script and register. About 68% of texts are in Indian languages or Hinglish. Published as arunp0/laya-indic-synthetic; see SYNTHETIC_DATA.md.
• Real-text questions (v2 and v3): 1–3 sentence snippets cut from Sangraha "verified" documents (AI4Bharat, CC-BY-4.0). The generator picks a topic from a fixed 16-topic list and writes 2–3 snippet-specific questions; topic options are assembled in code, and in 7% of cases the correct topic is removed so "none of the options" is the answer. The v2 part is published as arunp0/laya-indic-real-text. The v3 part covers all 11 languages roughly equally, with extra Odia (2,324 texts) and Urdu (2,212).
• Cause/effect questions (v3): Gemma reads a real snippet and writes a premise, a "what caused this?" or "what happened next?" question and two alternatives, in the snippet's language. Hindi, Marathi, Bengali, Tamil, Kannada, Malayalam, Punjabi, Odia and Urdu; not Telugu or Gujarati. Only 58% of generated questions passed verification.
• Passage yes/no questions (v3): a longer real snippet is the passage; Gemma writes yes/no questions about it in the passage language, about half answered "no". Gujarati, Tamil, Telugu, Bengali, Marathi and Kannada; not Hindi.

Limitations

• Text classification only. It does not do number, excerpt or citation questions, open-ended generation or multi-step reasoning. IndicCOPA (0.62) shows commonsense causal reasoning is still weak.
• Instructions in English. Questions and option descriptions should be in English; the text can be in any supported language.
• Long texts are truncated at 512 tokens per question (v2 allowed 1,024). XLM-R's tokenizer is 25–40% more compact on Indian scripts than v2's, which offsets part of this.
• Calibration is good on average but not everywhere. IndicCOPA and BoolQ-hi are poorly calibrated (ECE above 0.2). Recalibrate on your own data before using probabilities as thresholds.
• About 52% of the training labels come from language models (Qwen3, Gemma 4) and were checked only by another language model. Synthetic Indic text is cleaner and more formal than real user text. Sangraha is mostly news and web prose, not chat.
• Content warning. The safety and toxicity datasets (Aegis 2.0, Nemotron Safety Guard, Measuring Hate Speech, Civil Comments, Aya Red-teaming) contain offensive and harmful text. The model can classify such content but is not a tested content-moderation system.

FAQ

What is the best zero-shot classifier for Hindi and Hinglish?

In our tests Yama is the most accurate open zero-shot classifier for Indian languages. On 22 Indian-language tasks that no model was trained on (SIB-200 topic, IndicSentiment, IndicCOPA), it scores 0.813 accuracy, against 0.753 for bge-m3-zeroshot-v2.0, 0.586 for xlm-roberta-large-xnli, 0.532 for mDeBERTa-v3-base-xnli and 0.347 for bart-large-mnli. On Hinglish voice-assistant requests (Hinglish-TOP) it scores 0.942, but it was trained on that dataset's training split. Test it on a sample of your own data before relying on it.

Which languages does Yama support?

Instructions are written in English; the text can be in Hindi (Devanagari or romanised), Hinglish, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Urdu, Odia or English. It was also trained on Russian, Turkish, Azerbaijani, Kazakh, Uzbek, Georgian, Persian, Arabic, German, French, Spanish, Chinese, Japanese and Korean, and inherits XLM-RoBERTa's coverage of about 100 languages, but it is tested most on the languages above.

How is Yama different from bart-large-mnli or other NLI zero-shot models?

NLI models score each label separately as "does this text entail this text is about X?". Yama reads the question and all the options together and was trained directly on typed decisions, so it also answers yes/no questions and ordinal ratings, not just pick-one-label, and its probabilities are calibrated. See the comparison table under Benchmarks.

How is it different from asking an LLM?

Yama does not generate text. One forward pass of a 586M encoder returns probabilities for every option, so there is no output parsing, no sampling and no prompt-format drift. A single decision takes about 17 ms on an NVIDIA H100 and about 120 ms on 8 CPU threads (AMD EPYC); batched, it handles about 1,400 decisions per second on one H100. It is not a replacement for an LLM on open-ended or multi-step reasoning tasks.

Can I use it commercially?

Yes. The weights are Apache-2.0. All training data is CC0, CC-BY, Apache-2.0 or MIT, with no ShareAlike or NonCommercial sources and no proprietary-model outputs; see LICENSES.md for attributions.

How do I make it more accurate on my task?

Write clear option descriptions (they matter more than the keys), keep questions in English, and recalibrate the probabilities on a few hundred of your own labelled examples before using them as thresholds.

What happened to laya-indic?

This is the same repository, renamed with v3. arunp0/laya-indic redirects here, and the v1 and v2 weights are kept at revisions v1.0 and v2.0.

Acknowledgements and licence

• Architecture, training objective and inference library: THX-01 / thx01 by HAL-X AI (Apache-2.0).
• Benchmark harness: THX-01's bench/tasks.py, including its MASSIVE scenario descriptions and label humaniser, which were also used in some training instructions.
• Encoder: XLM-RoBERTa large, by Facebook AI (MIT). v1 and v2 used mmBERT-base, by JHU CLSP (MIT).
• Indian-language corpus: Sangraha, by AI4Bharat (CC-BY-4.0), source of the real-text questions.
• Label models: Qwen3 and Qwen3.6 (Alibaba Qwen team) and Gemma 4 (Google DeepMind), all Apache-2.0.

Copyright 2026 Arun Prabhakaran. The model is released under the Apache License 2.0. See NOTICE for the required attributions and LICENSES.md for every dataset's authors and licence.

Citation

@misc{prabhakaran2026yama,
  title  = {Yama: a zero-shot typed-decision model for Indian languages and Hinglish},
  author = {Prabhakaran, Arun},
  year   = {2026},
  url    = {https://huggingface.co/arunp0/yama-indic}
}